The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes a local bash script (scripts/systemone.sh) to interact with decision model APIs. The script employs security best practices, such as umask 077 to restrict permissions on temporary files and automated cleanup of session data using traps.
[INDIRECT_PROMPT_INJECTION]: The skill manages an attack surface where untrusted data (e.g., user-provided text) is ingested into the state field of API requests. It documents the risk of model steering and provides explicit mitigations.
Ingestion points: The state field in JSON requests processed by scripts/systemone.sh and defined in SKILL.md.
Boundary markers: The skill recommends using backticks in references/question-design.md to delimit state paths and reduce instruction confusion.
Capability inventory: The skill script performs network requests (curl) to decision model endpoints.
Sanitization: The skill lacks automated sanitization in the script logic, relying instead on user-defined prompt engineering, instruction tightening, and confidence gating described in the documentation.
[DATA_EXFILTRATION]: The helper script transmits the TYPESAFE_API_KEY to configured endpoints. It defaults to well-known or local services (such as api.typesafe.ai and 127.0.0.1) and provides clear instructions to avoid secret exposure in logs, code, or request files.

system-one